From GWAS Hits to Drug Targets: A Variant Prioritization Workflow

Q: How many genes should I expect from a typical GWAS locus?

The L2G model assigns high scores (>= 0.5) to a median of one gene per locus, but 10–20% of loci have two or more genes with L2G >= 0.5. Multi-signal loci (loci with two or more independently associated variants identified by GCTA-COJO) are more likely to have multiple causal genes than single-signal loci.

For Research Use Only. Not for use in diagnostic procedures or clinical decision-making.

A GWAS identifies genomic regions associated with a trait or disease. But a Manhattan plot peak is not a drug target — it is a pointer spanning dozens to hundreds of correlated variants, most of which are non-coding, none of which are guaranteed to be causal, and few of which act through the nearest gene. Turning that peak into a testable target hypothesis requires a systematic prioritization workflow: fine-mapping to isolate credible causal variants, variant-to-gene mapping using functional genomics evidence, multi-evidence integration into a ranked gene list, and target assessment for tractability and safety.

This guide walks through each step of the post-GWAS variant prioritization pipeline — from fine-mapping through drug target validation — with the tools, data sources, and decision rules that distinguish a well-prioritized target from a false lead. It is written for research teams that have GWAS summary statistics in hand and need to translate statistical signals into testable drug target hypotheses.

Infographic showing fine-mapping: a GWAS Manhattan plot with a significant peak zoomed into a regional plot highlighting multiple SNPs in LD, then narrowed to a 95% credible set of causal variants Figure 1: Fine-mapping narrows a GWAS locus from hundreds of correlated variants to a 95% credible set — the subset of variants that together have a 95% probability of containing the causal variant.

Why GWAS Hits Are Not Drug Targets

A GWAS identifies a statistical association between a genetic variant and a phenotype. It does not identify a causal gene, a direction of effect, or a therapeutic mechanism. Between a significant GWAS peak and a tractable drug target lie at least six layers of uncertainty:

The causal variant is usually not the lead SNP. The variant with the smallest p-value is often just the best-genotyped proxy in a block of highly correlated variants (LD r² > 0.8). The true causal variant may be ungenotyped, rare, or structural.
The causal gene is often not the nearest gene. Approximately 50% of GWAS signals map to genes other than the closest one. Regulatory elements can loop over intervening genes to contact promoters hundreds of kilobases away.
Most GWAS variants are non-coding. Roughly 90% of GWAS-associated variants lie outside protein-coding exons, in enhancers, promoters, insulators, or unannotated regions. Their functional effects are mediated through gene regulation, not protein sequence alteration.
Pleiotropy is common. Many GWAS loci influence multiple traits through the same or different genes. A variant that associates with both disease risk and a biomarker may act through the biomarker pathway — or through an unrelated mechanism.
Effect direction matters for drug development. A target where loss-of-function reduces disease risk suggests an inhibitor; a target where gain-of-function is protective suggests an activator. Without direction-of-effect evidence, the therapeutic modality is unspecified.
Population context shapes effect. Variant effect sizes, allele frequencies, and LD patterns differ across populations. A prioritization valid in one ancestry group may not transfer to another without explicit cross-population validation.

Each subsequent section of this guide addresses one or more of these uncertainty layers, building toward a ranked, evidence-weighted list of candidate drug targets.

For projects starting from raw variant calls rather than GWAS summary statistics, variant calling quality and GWAS analysis with appropriate covariate adjustment are prerequisites to the prioritization workflow described here. Cross-population validation also requires careful adjustment for ancestry — population structure analysis provides the principal component and ancestry estimation framework needed to control for population stratification in multi-ancestry GWAS and downstream prioritization.

Fine-Map to Credible Sets

The first priority after identifying significant GWAS loci is to narrow each locus to a manageable set of candidate causal variants. This process — statistical fine-mapping — uses the pattern of association signals and LD structure to compute, for each variant, the posterior probability that it is causal.

When a Locus Has One Causal Signal

For loci with a single independent association signal, two Bayesian methods dominate:

SuSiE (Sum of Single Effects). Decomposes the regression coefficients as a sum of "single-effect" vectors, each with exactly one non-zero element. SuSiE directly produces credible sets — subsets of variants that collectively have a specified probability (typically 95%) of containing the true causal variant. It handles high LD (r² > 0.99) naturally and outperforms stepwise conditional analysis when multiple causal variants are present. Available as the susieR R package.
FINEMAP. Uses a stochastic shotgun search to explore the space of causal configurations efficiently. Well-suited to loci with a single causal variant and moderate numbers of SNPs (up to ~5000 per locus). Produces posterior inclusion probabilities and Bayes factors per variant.

For a locus with one signal, both methods yield similar results. SuSiE is preferred when there is any suspicion of multiple independent signals at the same locus.

When a Locus Has Multiple Signals

For loci where conditional analysis (GCTA-COJO) identifies multiple independently associated variants, SuSiE can model them simultaneously by fitting multiple single-effect components. FINEMAP can also handle multiple causal variants, but the computational cost scales exponentially with the number of assumed causal variants.

The output from either method is a 95% credible set — typically 5 to 50 variants per signal — that becomes the input to functional annotation. Variants outside the credible set can be deprioritized regardless of their p-values.

Annotate Variant Function

Three-panel figure showing variant annotation: variant effect predictor classification table (missense, splice, regulatory, intergenic); regulatory overlap heatmap using ENCODE cCREs and tissue-specific chromatin states; LD proxy matrix showing correlated variants in different populations Figure 2: Variant annotation across three functional dimensions — predicted molecular consequence, regulatory overlap, and LD proxy identification across populations.

Once credible sets are defined, each variant in the set is annotated across three functional dimensions to distinguish likely causal variants from passengers.

Molecular Consequence

The Variant Effect Predictor (VEP) from Ensembl classifies each variant by its most severe molecular consequence: missense, splice donor/acceptor, stop gained/lost, frameshift, inframe indel, synonymous, 5'/3' UTR, intronic, intergenic. Coding variants, particularly those predicted damaging by Polyphen, SIFT, or CADD, are prioritized — but they account for fewer than 5% of credible set variants in most GWAS loci. The remaining >95% require regulatory annotation.

Regulatory Overlap

Overlap each variant with regulatory element catalogs:

ENCODE cCREs (candidate cis-regulatory elements): promoter-like, enhancer-like, and CTCF-only elements across hundreds of human cell types and tissues.
Tissue-specific chromatin states from Roadmap Epigenomics or ENCODE: active TSS, flanking TSS, transcribed, strong/weak enhancer, bivalent, heterochromatin.
Transcription factor binding sites (TFBS) from ChIP-seq peaks in relevant tissues (ENCODE, ReMap).

A non-coding variant inside an enhancer active in disease-relevant tissue is a stronger candidate than an intergenic variant in a repressed region, even if both have identical fine-mapping posterior probabilities.

LD Proxy Identification

The lead variant in the credible set may be a proxy for a rarer or ungenotyped causal variant. Use LDlink or PLINK with a population-matched reference panel (1000G, TOPMed, gnomAD) to identify all variants in high LD (r² > 0.8) with each credible set variant. These proxies should also be annotated — the true causal variant may be among them.

Map Variant to Gene

This is the most challenging step. A non-coding variant can regulate a gene 500 kb away while ignoring the three genes that are physically closer. Four orthogonal evidence types connect variants to their target genes:

eQTL and pQTL Colocalization

If the GWAS variant (or its LD proxy) also associates with expression of a nearby gene in a relevant tissue (eQTL) or with protein levels (pQTL), the gene is a strong candidate. Colocalization tests whether the GWAS signal and the eQTL/pQTL signal at the same locus share a causal variant — i.e., whether the same SNP drives both the disease association and the expression/protein change.

COLOC: Bayesian colocalization using summary statistics. Computes posterior probabilities for five hypotheses, including H4 (both traits share a causal variant). PP.H4 > 0.75 is the standard threshold for colocalization.
eCAVIAR: Extends the COLOC framework to allow multiple causal variants per locus.
Data sources: GTEx v8 (49 tissues), eQTL Catalogue (126 harmonized datasets), UKB-PPP (plasma pQTLs for ~3000 proteins).

A GWAS variant that colococalizes with a liver eQTL for a drug-metabolizing enzyme (PP.H4 > 0.8) provides strong evidence that the enzyme is the causal gene at that locus.

Chromatin Interaction

Promoter capture Hi-C and HiChIP data map physical contacts between regulatory elements and gene promoters in three-dimensional nuclear space. An enhancer containing a credible set variant that physically contacts the promoter of Gene A, and does not contact Gene B or Gene C, points to Gene A as the target — regardless of linear genomic distance.

Data sources: Jung et al. (2019) promoter capture Hi-C in 27 human cell types (available through Open Targets), 3D Genome Browser, ENCODE Hi-C.

Variant Effect Prediction on Regulatory Function

Deep learning models trained on large-scale functional genomics data can predict the regulatory impact of non-coding variants:

Enformer (DeepMind): Predicts gene expression and chromatin states from 200 kb of DNA sequence context.
DeepSEA: Predicts chromatin effects (TF binding, DNase sensitivity, histone marks) of non-coding variants.
ABC (Activity-by-Contact) model: Combines enhancer activity (ATAC-seq/DNase-seq/H3K27ac) with Hi-C contact frequency to estimate the quantitative effect of an enhancer on each gene's expression.

Rare Variant Evidence

Rare coding variants provide orthogonal support. If rare loss-of-function mutations in a gene produce a phenotype that aligns with the GWAS trait — for example, rare PCSK9 loss-of-function variants lowering LDL cholesterol, matching the GWAS signal at the PCSK9 locus — that gene is among the strongest possible drug targets. Gene-based rare variant burden tests from Genebass (UK Biobank exomes, n ≈ 400,000) and gnomAD provide per-gene association statistics for thousands of traits.

Integrate Evidence and Score Genes

At this point, you have multiple evidence layers pointing from variants to genes. The question is how to combine them into a single ranked list.

The Minimum-Rank Integration Strategy

A systematic benchmark by Moix et al. (2025, bioRxiv) compared five evidence integration strategies across 30 complex traits, combining GWAS scores, eQTL Mendelian randomization, pQTL Mendelian randomization, and rare variant burden data. The winner: minimum rank.

For each gene, compute its rank (1 = best) according to each individual evidence type. The integrated score is the minimum (best) of these ranks. This approach works because orthogonal evidence types have different tissue/cell-type coverage, different power to detect effects at different allele frequencies, and different sensitivity to confounding. A gene strongly supported by any single method — even if invisible to other methods — is a genuine candidate.

Weighted-sum approaches performed worse because they penalize genes that are inaccessible to certain methods (e.g., a gene without a measured plasma pQTL) — effectively punishing biological reality with low scores.

Open Targets Locus-to-Gene (L2G) Score

The Open Targets Platform provides a pre-computed L2G machine learning score for ~350,000 gene–locus pairs. The L2G model — an XGBoost classifier trained on gold-standard positive and negative gene sets — integrates:

Fine-mapping-weighted genomic distance
eQTL colocalization (126 tissue/cell types)
pQTL colocalization (UKB-PPP)
Promoter capture Hi-C (27 cell types)
Enhancer–TSS correlation
VEP functional consequence

The output is a calibrated score from 0 to 1, where ~80% of genes scoring above 0.8 are gold-standard positives. L2G scores serve as a strong default prioritization when custom multi-evidence integration is not feasible.

For research teams requiring custom scoring pipelines, variant discovery and drug target identification integrates fine-mapping, variant annotation, V2G mapping, and evidence scoring into a single analytical workflow.

Assess Target Tractability and Safety

A gene may be the causal gene at a GWAS locus with airtight multi-evidence support — and still be a poor drug target. The final prioritization step evaluates whether the gene encodes a protein that can be modulated by a drug.

Table 1: Drug Target Tractability and Safety Assessment Criteria

Assessment Criterion	What to Check	Key Resources
Protein class druggability	Is the encoded protein a member of a traditionally druggable family (GPCR, kinase, ion channel, nuclear receptor, protease, transporter)?	Open Targets Tractability, Pharos, DrugBank
Structural druggability	Does the protein have a binding pocket predicted to bind drug-like small molecules with high affinity?	PDB, AlphaFold structures, SIFT, DoGSite
Clinical precedence	Are there approved drugs or clinical-stage compounds targeting this protein for any indication?	ChEMBL, ClinicalTrials.gov, Open Targets
Safety liability	Do genetic or pharmacological data suggest on-target toxicity? Are there adverse event signals from known modulators?	SIDER, FDA Adverse Event Reporting System (FAERS), gnomAD constraint metrics
Assay tractability	Can target engagement be measured in a biochemical or cell-based assay suitable for high-throughput screening?	PubChem BioAssay, DrugBank, literature
Direction of effect match	Does the therapeutic hypothesis (inhibition vs. activation) match the genetic direction of effect? Loss-of-function protective → inhibitor; gain-of-function protective → activator.	GPS-DOE, Genebass, ClinVar

The Open Targets Platform provides pre-computed tractability assessments across small molecule, antibody, and PROTAC modalities, updated quarterly. Targets in the top quartile of both genetic evidence and tractability scores are strong candidates for entry into a drug discovery pipeline.

Build a Validation Plan

Four-panel validation plan infographic: orthogonal replication in independent cohorts, CRISPR perturbation of enhancer sequence, phenotype rescue in disease model, and human genetics replication across multiple evidence types Figure 3: A defensible validation plan spans orthogonal replication, molecular mechanism confirmation, phenotype rescue, and human genetics replication — each layer strengthening the target hypothesis.

Prioritization produces a hypothesis. Validation is what separates a computational prediction from a drug development program. A defensible validation plan includes:

Orthogonal replication. Does the variant–gene–phenotype association replicate in an independent cohort with a different genotyping platform and ancestry composition? A prioritization that holds in both European and East Asian cohorts is more robust than one detectable only in the discovery dataset.
Molecular mechanism. For a non-coding variant predicted to disrupt an enhancer: does CRISPR perturbation (CRISPRi/CRISPRa) of the enhancer sequence alter target gene expression in a disease-relevant cell type? For a coding variant: does overexpression or knockout of the variant allele alter the cellular phenotype in the expected direction?
Phenotype rescue. If the target gene is modulated — by siRNA, CRISPR, or a tool compound — in a disease-relevant cell or animal model, does the disease-associated phenotype normalize? This is the strongest pre-clinical evidence that the target is worth prosecuting.
Human genetics replication. Confirm that rare variant burden, Mendelian randomization, or family-based studies in independent cohorts support the same gene–phenotype connection. A target supported by both common-variant GWAS and rare-variant burden evidence is substantially more credible than one supported by GWAS alone.

For researchers managing a portfolio of GWAS loci across multiple traits, multi-omics integration connects variant-level associations with expression, methylation, and protein-level QTL data to narrow the mechanistic gap between association and function. For a detailed walkthrough of the colocalization and mediation analysis methods that power this integration, see the guide on multi-omics QTL integration.

Frequently Asked Questions

What is the difference between fine-mapping and clumping?

Clumping (PLINK --clump) selects the variant with the smallest p-value in each LD block and discards the rest. It does not estimate which variant is causal — it simply picks a representative. Fine-mapping computes posterior inclusion probabilities for each variant in the block, accounting for the correlation structure, and produces a credible set. For drug target prioritization, clumping is insufficient because it can select a non-causal proxy as the "representative" without quantifying uncertainty. Fine-mapping is required.

How many genes should I expect from a typical GWAS locus?

The L2G model assigns high scores (≥0.5) to a median of one gene per locus, but 10–20% of loci have two or more genes with L2G ≥ 0.5. Multi-signal loci (loci with two or more independently associated variants identified by GCTA-COJO) are more likely to have multiple causal genes than single-signal loci.

How do I handle pleiotropy — when one variant associates with multiple traits?

Pleiotropy is not a problem to be eliminated; it is a feature that can strengthen target prioritization. If a credible set variant associates with both disease risk and a biomarker in the expected direction, the mechanistic link is stronger. If the variant associates with unrelated traits through different genes at the same locus, those genes may need to be de-prioritized for the specific disease target program. Cross-trait colocalization and the PIPE framework are designed to leverage pleiotropy rather than be confounded by it.

What should I do if my GWAS locus has no credible eQTL or chromatin interaction evidence?

Absence of functional evidence is not evidence of absence. Some causal variants act through mechanisms not captured by current functional genomics databases: splicing regulation in a tissue not profiled by GTEx, cell-type-specific effects masked in bulk-tissue eQTL data, temporal effects during development, or effects on non-coding RNA genes. Document what was tested and found absent, and flag the target as having lower evidence confidence — but do not discard it solely on the basis of missing functional annotation.

References:

Wang G, Sarkar A, Carbonetto P, Stephens M. A simple new approach to variable selection in regression, with application to genetic fine mapping. Journal of the Royal Statistical Society, Series B: Statistical Methodology. 2020;82(5):1273–1300. doi:10.1111/rssb.12388
McDonagh EM, Trynka G, McCarthy M, et al. Human genetics and genomics for drug target identification and prioritization: Open Targets' perspective. Annual Review of Biomedical Data Science. 2024;7:59–81. doi:10.1146/annurev-biodatasci-102523-103838
Chen R, Duffy A, Petrazzini BO, et al. Expanding drug targets for 112 chronic diseases using a machine learning-assisted genetic priority score. Nature Communications. 2024;15:8891. doi:10.1038/s41467-024-53333-y
Mountjoy E, Schmidt EM, Carmona M, et al. An open approach to systematically prioritize causal variants and genes at all published human GWAS trait-associated loci. Nature Genetics. 2021;53(11):1527–1533. doi:10.1038/s41588-021-00945-5
Finan C, Gaulton A, Kruger FA, et al. The druggable genome and support for target identification and validation in drug development. Science Translational Medicine. 2017;9(383):eaag1166. doi:10.1126/scitranslmed.aag1166
Moix S, Sadler MC, Kutalik Z. Integration of genetic evidence to identify approved drug targets. bioRxiv. 2025. doi:10.1101/2025.10.10.681636

For Research Use Only. Not for use in diagnostic procedures or clinical decision-making.

* Designed for biological research and industrial applications, not intended for individual clinical or medical purposes.